Solving and bounding Stein equations

Ferdinand Rapin and Yvik Swan, Representations of solutions to Stein equations and their derivatives (2026). arXiv:2609.17881 · Article PDF.

The method

For a density pp on (ℓ,u)(\ell,u), choose a nonvanishing weight ww and set γ=(wp)′/p\gamma=(wp)'/p. The first-order Stein equation is

w(x)fh′(x)+γ(x)fh(x)=h(x)−E[h(Z)].w(x)f'_h(x)+\gamma(x)f_h(x)=h(x)-\mathbb E[h(Z)].

With the appropriate boundary condition, integration gives the canonical solution

fh(x)=∫ℓx(h(t)−E[h(Z)])p(t) dtw(x)p(x).f_h(x)=\frac{\displaystyle\int_\ell^x\bigl(h(t)-\mathbb E[h(Z)]\bigr)p(t)\,dt}{w(x)p(x)}.

The method represents this solution and its derivatives through an array of signed kernels. Differentiation and integration identities move between derivative orders of the solution and of the test function. The choice of weight determines which correction terms cancel; in general these terms must be retained.

Controlling the test function and the absolute kernel integrals gives pointwise envelopes. Taking a supremum over xx gives uniform Stein factors; multiplying by a chosen weight gives weighted factors. Kernel signs can identify extremising tests and establish sharpness. Each bound depends on the test class, derivative order, moment assumptions, and boundary conditions. See the article for the representations and their hypotheses.

Companion papers and addendum

The following September 2026 preprints organise the bounds by test class: Kolmogorov, bounded measurable, Wasserstein, Fortet–Mourier, and integer-order Zolotarev. They distinguish exact envelopes, explicit upper bounds, and derivative orders for which the requested bound is unavailable.

Related illustrations: solving Stein equations with kernels.

Back to contents